The question everyone is asking this week #
Apple’s M5 Ultra Mac Studio (announced August 25, 2026) finally gives local-AI buyers a single box with up to 512GB of unified memory at 1.2TB/s. But it arrived into a market where the cluster path already existed: four NVIDIA DGX Sparks (4x 128GB) or four AMD Ryzen AI Halo boxes (4x 128GB) pool the same 512GB over the network. Every community thread this week is the same fight: one big box, or four small ones?
The three paths, at currently verified prices:
| 1x Mac Studio M5 Ultra | 4x NVIDIA DGX Spark | 4x AMD Ryzen AI Halo | |
|---|---|---|---|
| Total memory | |||
| 256GB (verified) / 512GB (late Oct) | 512GB | 512GB | |
| Price | |||
| $10,799 (256GB) / 512GB price TBA | $18,796 ($4,699 x 4) | $15,996 ($3,999 x 4) | |
| Memory bandwidth | |||
| 1,200 GB/s, one pool | 4 x 273 GB/s, split across nodes | 4 x 256 GB/s, split across nodes | |
| Software | |||
| MLX, llama.cpp, Metal | CUDA, TensorRT-LLM (cloud parity) | ROCm / Vulkan | |
| Wall power (load) | |||
| ~300-400W, one outlet | ~960W+ across four outlets | ~480W across four outlets | |
| Failure mode | |||
| One box down = zero capacity | One box down = -25% memory pool | One box down = -25% memory pool | |
| Ships | |||
| Sept 22 (512GB late Oct) | In market since 2025 | In market since 2026 |
The value play we also track: four GMKtec EVO-X2 boxes - the same Strix Halo silicon as the Halo dev kit - cost $7,996 ($1,999 x 4) for the same 512GB. More on why that matters below.
The bandwidth math that decides decode speed #
Token generation is memory-bandwidth-bound: every token reads the active model weights from memory, so decode speed tracks bandwidth divided by bytes-read-per-token. Take GLM-5.2 at Q4 - roughly 20GB of active experts per token - and compute the theoretical decode ceiling for each single memory pool:
| Memory pool | Bandwidth | Ceiling (GLM-5.2 Q4, ~20GB active) |
|---|---|---|
| M5 Ultra (one box) | 1,200 GB/s | ~60 tok/s |
| DGX Spark (one box) | 273 GB/s | ~13 tok/s |
| Ryzen AI Halo (one box) | 256 GB/s | ~12 tok/s |
These are ceilings, not benchmarks - real-world decode lands at 50-70% of the formula depending on runtime and quantization. But the ratio is what matters: the M5 Ultra’s single pool moves data roughly 4.4x faster than any one cluster node.
The cluster tax #
“But four Sparks have 1,092 GB/s aggregate!” True, and mostly irrelevant: cluster nodes cannot pool bandwidth for one model the way UltraFusion pools dies inside one package. When you shard a model across four boxes, every generated token requires activations to cross the network between layers. The interconnect (Thunderbolt 5, ConnectX-7, or Ethernet) adds latency to every single token, so clustered decode lands somewhere between the fastest single node and a naive 4x - community multi-Spark runs typically see 1.5-2.5x single-node throughput, not 4x. The EXO project’s published Spark clustering work is the best public reference here; treat any “linear scaling” claim from a vendor diagram as marketing until you see the harness.
Where the cluster genuinely wins is prefill and throughput-heavy batch work. Prompt processing is compute-bound, four Blackwell GPUs have serious aggregate compute, and independent requests can be routed to different nodes with real linear scaling. If you serve many users or process very long prompts, the calculus shifts toward the cluster.
What each path is actually for #
Buy the M5 Ultra if you want the fastest single-stream decode per dollar and one machine to maintain. The verified 256GB config at $10,799 already runs GLM-5.2, DeepSeek V4, or Kimi-class models at Q4 with room for KV cache; the 512GB config (price unpublished, ships late October) is the only consumer path to full-precision frontier-class open weights in one box. MLX support for new open-weight drops is usually same-week.
Buy the DGX Spark cluster if your stack is CUDA-shaped. TensorRT-LLM, vLLM, and every cloud-NVIDIA workflow runs unmodified, prefill throughput is the best of the three, and what you learn transfers directly to rented A100/H100 time. You are paying a ~$8,000 premium over the Halo cluster for that software ecosystem and NVIDIA’s reference-platform polish.
Buy the Ryzen AI Halo cluster if budget rules everything and you accept the riskiest software path. Four EVO-X2s at $7,996 is the cheapest 512GB on the market by a wide margin - less than one M5 Ultra 256GB - but ROCm/Vulkan support for large MoE models lags, and you are the beta tester. For 30B-class workloads it is genuinely good; for frontier MoE models, expect Vulkan-flag weekends.
Honest caveats #
The M5 Ultra 512GB price is not public. Apple ships that config late October 2026; everything in this guide prices the verified $10,799 256GB config and treats 512GB pricing as unknown until Apple lists it. Apple’s June 2026 price hikes (RAM-driven, industry-wide) make a big 512GB premium likely - do not budget from the old M3 Ultra upgrade math ($4,000 for 96GB to 512GB in 2025). - Cluster throughput numbers above are community-reported, not ours. The 1.5-2.5x figure varies hugely by model, sharding strategy, and interconnect. If clustered inference is your plan, benchmark your exact model with EXO or llama.cpp RPC before buying four of anything. - Power is a real cost. Four Sparks under sustained load draw roughly twice what four Halos do and around 2.5x one M5 Ultra; at US average rates a 24/7 inference node adds $800-1,000/year per kW. - Prices move. DRAM shortages drove Apple’s 2026 price hikes and NVIDIA’s Spark increase ($3,999 to $4,699 in February 2026). Verify every price at purchase time.
The one-line verdict #
For one person running one big model: M5 Ultra - no cluster tax, the fastest decode, one power outlet. For a CUDA team or a multi-user service: the Spark cluster. For maximum memory per dollar with tolerance for pain: four EVO-X2s. And if you are between budgets, the 256GB M5 Ultra at $10,799 is the verified sweet spot - it runs nearly everything the 512GB tier does at Q4, today, at a known price. Check what a specific rig can run with the rig finder, compare live tok/s submissions on the hardware pages, and see our guide on running GLM-5.2 locally for the model-side math.